A Novel Entropy Based Algorithm to Remove Silence from Speech andClassifying the Residue as Voiced/unvoiced Regions

摘要

For any speech synthesis, voiced portion of speech plays a crucial role. Major researchers have focussed on the most sophisticated statistical approaches whereas least importance was given to the time-domain or frequency domain approaches citing their limitations. The issues of statistical approaches are dealt with by adding new features making them more complex rather than resolving complexity. So we propose an algorithm which uses one feature namely sample entropy to classify speech signal. In our proposed algorithm, silence removal is achieved by fuzzy entropy and sample entropy is used to classify the residual speech signal as voiced or unvoiced regions. The performance of the proposed algorithm is analysed using TIMIT database. The proposal outperforms the existing approaches with a 94.98 % accuracy of information during silence removal from speech signals and the classification rate is analysed using Receiver Operating Characteristics (ROC) which yields an accuracy of 92.78 %.